call_runner rejects a top-level JSON array, so a runner cannot pass one through

call_runner decides how to hand back a response from its Content-Type alone: JSON is parsed into result.data, anything else arrives unparsed in result.content. A body that is valid JSON but not an object falls between the two and raises instead:

livepeer_gateway.errors.LivepeerGatewayError: Live runner call expected JSON object, got list

Repro

import asyncio
from aiohttp import web
from livepeer_gateway.live_runner import call_runner


async def main():
    async def handler(request):
        return web.json_response([{"label": "llama", "score": 0.99}])

    app = web.Application()
    app.router.add_post("/call", handler)
    runner = web.AppRunner(app)
    await runner.setup()
    site = web.TCPSite(runner, "127.0.0.1", 0)
    await site.start()
    port = site._server.sockets[0].getsockname()[1]
    try:
        result = await call_runner(f"http://127.0.0.1:{port}/call", payload={"inputs": "x"})
        print("data:", result.data, "content:", result.content)
    finally:
        await runner.cleanup()


asyncio.run(main())

Why it matters

A runner that proxies somebody else's API does not choose its response shape, and a top-level array is a common one. Every Hugging Face inference task except text-to-image answers with one: classification, object detection, embeddings, sentiment. Such a runner is currently unreachable through the SDK.

The workaround available to a proxy runner is to make the response stop claiming to be JSON so it takes the opaque-bytes path. In runner-app-examples#82 that is nginx relabelling a JSON body as text/plain, which is a lie told to route around a type check, and it is the only lever config has, since nginx can pin, inject, and relabel but cannot rewrite a body.

Proposal

Treat non-object JSON the way ndjson and binary are already treated: return the body unparsed in result.content with result.content_type intact, leaving result.data as {}.

Objects keep today's behavior exactly, including session_id extraction, so no working call changes: the only path affected is the one that raises today. content_type still reports application/json, so a caller can tell what it is holding and json.loads it.

The current strictness is deliberate (tests/test_call_runner_raw.py::test_json_array_still_rejected), and it is right for the control-plane calls that read protocol fields out of the body, such as proxy create and trickle channel remove. Those stay strict. It is the pass-through data path where an array is data rather than a malformed reply.

An alternative is to keep data dict-only and expose the parsed value separately (result.json), but that adds API surface for something content already carries.

Comienza en livepeer_gateway/live_runner.py, en call_runner, y lee tests/test_call_runner_raw.py, especialmente test_json_array_still_rejected. Comprueba las rutas existentes de respuesta object, ndjson y binary, y después verifica que un array JSON de nivel superior se devuelva a través de content sin cambiar el manejo de object. Las pruebas específicas de raw-response deberían confirmar el nuevo comportamiento y conservar un manejo estricto del control-plane.